EP.02: The Growth Infrastructure

From Snap to the plumbing layer

Bryan Kim opens by introducing himself: he has been at a16z for about five and a half years, with a focus on growth and consumer. Before that he was at Snap from the time it was around 100 people and pre-revenue, and watched it grow to 4,000 people and $2 billion in revenue. Most importantly for this conversation, he built Snap’s growth team.

That experience taught him that the usual growth vocabulary of paid acquisition, retention, engagement and channel optimization is not the core issue. The core issue is whether you understand why people use your product, and whether they use, browse and buy the way you want them to. To understand that, he says, you have to master what he calls “the plumbing layer.” When the Hilbert founders first explained the company to him, his reaction was: “You’re plumbers.”

He describes three growth personas. There are the people who handle LTV/CAC and optimization, which he calls a tractable math problem. There are the creatives who think about channels and formats. And there are the plumbers. Bryan says he has interviewed hundreds of candidates for that role, because it requires systems-level thinking, fluency with the data schema and the ability to tie data to product insight.

Why the company is called Hilbert

Nazlı explains the name. She is a mathematician by education and has long been fascinated by the foundational crisis of mathematics, when contradictions and paradoxes appeared in the math people were using and a group of mathematicians argued for a solid foundation. In growth, she kept running into the same problems: unclear definitions and numbers that did not match. She wanted to build a foundation for the practice, so she borrowed the name from Hilbert’s program, David Hilbert’s attempt to address the crisis. The company started as Hilbert’s Program before becoming Hilbert.

Lessons from Getir

Nazlı joined Getir as a power user and started its growth function alongside a separate performance marketing organization. Last-mile delivery with a 15-minute promise brings real constraints: stores must be replenished and riders available, so you cannot simply push growth as hard as possible, and thin margins limit what you can spend on paid. That environment built her habit of aligning everyone around something meaningful.

At Getir, Nazlı focused on how to retain and monetize the users the paid team brought in, and on feeding that learning back to paid channels. Her mathematical background made her question standard definitions. Was 30 days without an order really churn? Was a customer with a high AOV really the most valuable, or might someone else have higher frequency or better profit? That led her to use data science as fully as possible. If you have the data and can predict, she asks, why rely on a human-made definition?

An opinionated infrastructure for growth

Asked for a one-liner, Nazlı describes Hilbert as “an opinionated infrastructure for growth.” Instead of bending the solution to fit each company, Hilbert maps each company onto a large, generalized schema with rich context. Deep learning models on that schema study the whole customer lifecycle and generate labels: predicted LTV from the very first order, a churn score adjusted to each customer’s frequency, risk and reactivation scores, and a set of propensity scores for personalization.

She calls it an executive kitchen for chefs. Growth, in her view, is a thinking job, closer to a mathematician or philosopher who looks at data, forms a hypothesis, and tests and learns quickly. Yet most growth people spend their time stitching data together. On top of the labels, Hilbert runs AI agents that know the schema and the scores and respect the deep learning definitions.

Whack-a-mole versus systems thinking

From the investor side, Bryan sees companies fix retention, then engagement, then AOV one by one, like whack-a-mole. Nazlı explains why that fails: the metrics are functions of each other because they all measure customer behavior, making growth a multivariate optimization problem. Teams that try to lower CAC in isolation can end up with low-quality customers whose problems only show up after six months. Retention teams may not know who is coming in the door, or understand the unit economics of the discounts they use. Monetization moves such as raising prices or fees change customer behavior and affect the acquisition engine, and those effects do not appear in a one-month A/B test. The result is local minima and maxima where metrics move but the company does not grow.

Her answer is a metric system that is itself non-gameable. If a company runs 50% off sitewide, revenue, customers and acquisition rise, but Hilbert’s labels will not move, because the history shows a sitewide discount is probably not an LTV-growing action. A shared metric system also lets teams keep themselves honest and work toward the same goal.

What getting it wrong costs

Asked what it costs when companies get this wrong, Nazlı says millions, and for some companies billions, especially in B2C, where paid marketing is part of life and older companies have “castles built on those channels” that cannot be switched off overnight. Add discounts and opportunity costs, and it becomes an expensive problem.

Why now, and who it serves

Bryan asks how Hilbert differs from AI-enabled dashboards and campaign tools, and why now. Nazlı says driving long-term growth takes more than piecemeal optimizations. Companies can pursue different north stars, such as profitability, aggressive growth or stability, and Hilbert is infrastructure that supports whichever one leadership chooses. Its backbone is deep learning and the metric system. LLMs and agents make that complexity conversational, which lets the whole company, not just the growth team, sit at the table.

She points to the back-and-forth between CMOs, growth teams and CFOs over defending budgets. Bryan shares a Snap story where an unexplained jump of 1 million DAU took 12 hours to trace to a country-level spike caused by IP spoofing, a question he says Hilbert could have answered right away.

Working with large enterprises

For large enterprises, Nazlı says the challenge is a big inactive user base, a large marketing budget and legacy definitions, which makes the work like fixing a plane while it is flying. Hilbert starts by deploying its labels, which often tell a different story about risk, value and preferences, then connects them back to the data warehouse so teams can activate them anywhere and tackle problems like discount dependency and diminishing returns on paid channels.